A Global View of How LLMs Value Human Time

Yifan He; Manon Reusens; Sofie Goethals; David Martens

cs.CYOct 13, 2026

A Global View of How LLMs Value Human Time

Authors: Yifan He, Manon Reusens, Sofie Goethals, David Martens

Abstract

Large language models (LLMs) increasingly act as personal assistants; however, little is known about how they capture national-level preferences on a global scale. Given that "time" is a universal and representative concept, we examine LLMs' time preferences by varying only users' country and measuring the compensation LLMs deem appropriate for a 60-minute appointment delay. We then compare LLMs' responses with GDP per capita on Purchasing Power Parity (GDP_PPP), Mean Income, and a literature-based index (Patience Score). Our results across 193 countries and 5 LLMs demonstrate that people's time across the world is evaluated in noticeably different ways among LLMs, as well as varying levels of robustness under prompt variations. LLMs consistently recommend the highest compensation for Europe, but mostly the lowest for Africa. Based on GDP_PPP, LLMs underestimate the value of time for users living in Asia and South America. These findings indicate that geographical patterns are embedded in LLMs and may translate in different treatment in agentic settings.

Figures & tables

Figure 1: Example of the derived compensation for a 60-minute delay in Claude-Haiku-4.5. This illustrates the differences in estimated time value across the globe.

Figure 2: Overview of the workflow. LLMs decide whether to accept a series of compensations for additional waiting. For each country, a transition value is determined at which LLMs flip answers using logistic regression. These numbers are visualised on world maps and compared against multiple indicators.

	Models	Africa (54)	Asia (46)	Europe (44)	N. America (23)	Oceania (14)	S. America (12)
Compensation (US$)	claude-haiku-4.5	13.1 $\pm$ 10.0	16.1 $\pm$ 8.3	20.7 $\pm$ 6.8	15.3 $\pm$ 9.0	15.1 $\pm$ 9.0	16.8 $\pm$ 8.2
	deepseek-v4-pro	2.3 $\pm$ 0.8	3.7 $\pm$ 0.9	6.6 $\pm$ 1.2	3.8 $\pm$ 0.9	4.3 $\pm$ 1.1	3.0 $\pm$ 0.8
	glm-5.2	1.1 $\pm$ 0.0	2.7 $\pm$ 0.1	6.3 $\pm$ 0.8	3.2 $\pm$ 0.4	2.6 $\pm$ 0.4	1.6 $\pm$ 0.3
	gpt-5.4-mini	4.6 $\pm$ 4.1	5.5 $\pm$ 5.5	8.8 $\pm$ 9.0	6.5 $\pm$ 6.4	5.6 $\pm$ 5.5	6.5 $\pm$ 6.0
	mistral-large-3	16.5 $\pm$ 7.5	16.2 $\pm$ 7.7	18.5 $\pm$ 6.0	17.5 $\pm$ 6.6	18.2 $\pm$ 6.2	15.4 $\pm$ 8.4
	Mean	7.5	8.8	12.2	9.3	9.2	8.7

Table 1: Accepted compensation (US$) per continent across three prompt variations.

Models	$R^{2}$ : GDP_PPP (179)	$R^{2}$ : Mean Income (113)	$R^{2}$ : Patience (111)
claude-haiku-4.5	$0.261\pm 0.285$	$0.264\pm 0.245$	$0.075\pm 0.071$
deepseek-v4-pro	$0.642\pm 0.046$	$0.816\pm 0.014$	$0.297\pm 0.022$
glm-5.2	$0.543\pm 0.034$	$0.839\pm 0.008$	$0.277\pm 0.021$
gpt-5.4-mini	$0.289\pm 0.240$	$0.392\pm 0.243$	$0.160\pm 0.102$
mistral-large-3	$0.091^{\dagger}\pm 0.035^{\dagger}$	$0.237^{\dagger}\pm 0.013^{\dagger}$	$0.067^{\dagger}\pm 0.030^{\dagger}$

Table 2: Linear Regression Results with three indicators: GDP_PPP, Mean Income, and Patience Score.

	Models	Africa (54)	Asia (46)	Europe (44)	N. America (23)	Oceania (14)	S. America (12)
Overvalued (%) to GDP_PPP	claude-haiku-4.5	62.2 $\pm$ 31.6	39.2 $\pm$ 25.0	52.8 $\pm$ 37.0	44.4 $\pm$ 36.1	61.9 $\pm$ 25.1	51.5 $\pm$ 18.9
	deepseek-v4-pro	53.8 $\pm$ 3.3	25.0 $\pm$ 4.3	51.2 $\pm$ 2.4	39.7 $\pm$ 12.0	83.3 $\pm$ 8.2	0.0 $\pm$ 0.0
	glm-5.2	53.2 $\pm$ 2.9	30.0 $\pm$ 6.6	56.9 $\pm$ 1.4	34.9 $\pm$ 9.9	69.0 $\pm$ 14.9	0.0 $\pm$ 0.0
	gpt-5.4-mini	44.2 $\pm$ 3.3	19.2 $\pm$ 15.3	56.1 $\pm$ 16.9	33.3 $\pm$ 16.5	40.5 $\pm$ 4.1	54.5 $\pm$ 15.7
	mistral-large-3	$51.0^{\dagger}\pm 1.3^{\dagger}$	$27.5^{\dagger}\pm 3.5^{\dagger}$	$45.1^{\dagger}\pm 1.7^{\dagger}$	$52.4^{\dagger}\pm 13.4^{\dagger}$	$67.9^{\dagger}\pm 15.2^{\dagger}$	$13.6^{\dagger}\pm 6.4^{\dagger}$
	Mean	52.9	28.2	52.4	40.9	64.5	23.9

Table 3: The percentage of overvalued countries by continent across three prompt variations.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: The count of responses under each reward quantity ($20, $30 and $40) across 10 representative countries. — Figure 3: The count of responses under each reward quantity ( $20,$ 30 and $40) across 10 representative countries.

LLMs	Country	Even (20)	Even (49)	Even (100)	Log (20)	Log (49)	Log (100)
claude-haiku-4.5	Afghanistan	24.9	23.5	>25.0	>25.0	>25.0	>25.0
	Switzerland	>25.0	>25.0	>25.0	>25.0	>25.0	>25.0
deepseek-v4-pro	Afghanistan	<1.0	1.8	<1.0	<1.0	1.6	<1.0
	Switzerland	9.7	9.8	10.0	8.5	9.9	9.8
glm-5.2	Afghanistan	<1.0	<1.0	<1.0	<1.0	<1.0	<1.0
	Switzerland	10.9	12.9	12.5	11.8	13.4	13.5

Appendix

Table 4: Results across six experimental settings, combining 20, 49, or 100 steps with either even or log spacing over the $1–25 compensation range.

Models	Checkpoints/Providers	Parameters (total / active)
Claude-Haiku-4.5	Official API:	Undisclosed
	claude-haiku-4-5-20251001
GPT-5.4-Mini	Official API:	Undisclosed
	gpt-5.4-mini-2026-03-17
DeepSeek-V4-Pro	OpenRouter:	1.6T / 49B
	streamlake/fp8

Appendix

Table 5: Models, API platforms, and parameter counts. We access fixed checkpoints when using the official API platform and fixed providers on OpenRouter. Parameter counts are taken from official model cards or checkpoint metadata; “Undisclosed” means the provider has not published the number.

Models	Cost (USD)
Claude-Haiku-4.5	11.4
GPT-5.4-Mini	9.6
DeepSeek-V4-Pro	17.8
GLM-5.2	28.4
Mistral-Large-3	2.2

Appendix

Table 6: Models and Cost in USD (Prompt 1).

Panel of Figure 4 — Figure 4: World maps showing the responded worth of time across five LLMs (in US $) under Prompt 1. Countries with transition values below$ 1 and above $25 are assigned the colors equals to$ 1 and $25, respectively.

Figure 5: Results across three prompt variations.

Panel of Figure 6 — Figure 6: Using China as an example, these figures show models’ responses across three prompt variations.

Panel of Figure 7 — Figure 7: Scatter plots and linear regressions of country-level compensation against the logarithm of GDP_PPP across the five LLMs (Prompt 1). Each dot represents a unique country, with colors indicating the continent to which it belongs.

Models	$R^{2}$ : GDP_PPP (83)	$R^{2}$ : Mean Income (83)	$R^{2}$ : Patience (83)
claude-haiku-4.5	$0.346\pm 0.314$	$0.213\pm 0.189$	$0.075\pm 0.075$
deepseek-v4-pro	$0.714\pm 0.039$	$0.796\pm 0.008$	$0.373\pm 0.027$
glm-5.2	$0.613\pm 0.040$	$0.815\pm 0.007$	$0.356\pm 0.018$
gpt-5.4-mini	$0.360\pm 0.296$	$0.360\pm 0.252$	$0.178\pm 0.133$
mistral-large-3	$0.375^{\dagger}\pm 0.050^{\dagger}$	$0.346^{\dagger}\pm 0.039^{\dagger}$	$0.102^{\dagger}\pm 0.026^{\dagger}$

Appendix

Table 7: Linear Regression Results with three indicators: GDP_PPP, Mean Income, and Patience Score (83 countries).

Jurisdictional Bias in Language Models

See all 9 papers

Need or Reputation? Region, National Income, and Academic Status in LLM Gatekeeping Decisions
Jun 27, 2026Nouar AlDahoul, Hezerul Abdul Karim, Myles Joshua Toledo TanSocial Bias in Language Models LLM Decision-Making
Computer Science Department, New York University Abu Dhabi UAE · Centre for Image and Vision Computing, Faculty of Artificial Intelligence and Engineering Multimedia University Malaysia · Department of Electrical and Computer Engineering, Herbert Wertheim College of Engineering University of Florida USA
Temporal Preference Concepts and their Functions in a Large Language Model
May 11, 2026Ian Rios-Sialer, Shantanu Darveshi, Shuai Jiang +4LLM Interpretability Temporal Reasoning in Language Models
AISC San Francisco, USA · AISC Munich, Germany · SPAR, AISC Berlin, Germany
Can LLMs Perceive Time? An Empirical Investigation
Mar 9, 2026Aniketh GarikaparthiTemporal Reasoning in Language Models LLM Inference
TCS Research

Jun 27, 2026J/K move · Enter open · S save

Need or Reputation? Region, National Income, and Academic Status in LLM Gatekeeping Decisions

Nouar AlDahoul, Hezerul Abdul Karim, Myles Joshua Toledo Tan

Computer Science Department, New York University Abu Dhabi UAE · Centre for Image and Vision Computing, Faculty of Artificial Intelligence and Engineering Multimedia University Malaysia · Department of Electrical and Computer Engineering, Herbert Wertheim College of Engineering University of Florida USA

Social Bias in Language Models LLM Decision-Making Language Model Bias Evaluation LLM Auditing

Open paper AskPDF arXiv

Equitable access to scientific knowledge often depends on informal gatekeeping decisions, particularly when resources such as paywalled articles, nonpublic datasets, or a researcher's CV must be shared selectively. We introduce a controlled simulation framework in which LLM-based professors must grant access to only one requester. Across prompts, requesters vary systematically by global region (Global North vs. Global South) and academic seniority, while all other factors remain constant. Most models prioritize PhD candidates, although preferences vary by model and scenario (e.g., Claude favors tenured professors when the request is for the professor's CV), and the smaller models' status choices are sensitive to the order in which options are presented. When the requesters' regions differ, the models split into two groups. The three large proprietary LLMs systematically favor requesters from the Global South, a pro-equity bias consistent with their safety alignment, whereas the two small open-weight models reverse this preference and favor the Global North, possibly reflecting weaker alignment or the geographic skew of their pre-training data. Linking choices to national income reveals that the large models' preference largely tracks economic need rather than the regional label: they strongly favor requesters from low-income countries, but their preference for the Global South is weak or absent when both countries have equal income, whereas the small models favor the Global North regardless of income. Because model scale, alignment, and training data are confounded in our sample, these explanations remain hypotheses for future work. Our findings show that LLMs embed differing normative assumptions about need and reputation that can shape gatekeeping decisions, which highlights the need to audit AI systems for fairness and value alignment before such tasks are delegated to them.

Abstract

Jurisdictional Bias in Language Models

Explore similar work

Need or Reputation? Region, National Income, and Academic Status in LLM Gatekeeping Decisions

Temporal Preference Concepts and their Functions in a Large Language Model

Can LLMs Perceive Time? An Empirical Investigation